Reproduce “Whole-Body Mobile Manipulation using Offline Reinforcement Learning on Sub-optimal Controllers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Snehal Jauhri, Vignesh Prasad, Georgia Chalvatzaki
The PEARL Lab at Technische Universität Darmstadt developed WHOLE-MoMa, a method for whole-body mobile manipulation of articulated objects that leverages sub-optimal classical controller data for offline reinforcement learning. This approach achieved 80% success on bimanual drawer tasks and 68% on complex cupboard opening and object placement tasks in direct sim-to-real transfer to a physical robot, without relying on teleoperated data.
Key insights
- Leveraging a classical Whole-Body Controller as a "structural prior" efficiently generates diverse, task-relevant data for offline RL, bypassing costly teleoperation data collection.
- Offline reinforcement learning can identify and "stitch together" improved, long-horizon behaviors from sub-optimal demonstrations, overcoming the myopia of classical controllers.
- Adapting IQL with "Q-chunking" and using expressive action-chunked Diffusion Policies enhances temporal consistency and expressivity for complex whole-body movements in high-dimensional spaces.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.